Self-Supervised Segmentation of River Scenes
Supreeth Achar, Bharath Sankaran, Stephen Nuske, Sebastian Scherer and Sanjiv Singh.
Abstract— Here we consider the problem of automatically
and accurately segmenting visual images taken from a boat
or low-flying aircraft. Such a capability is important for both
mapping rivers as well as following rivers autonomously. The
need for accurate segmentation in a wide variety of riverine
environments challenges the state of the art in vision-based
methods that have been used in more structured environments
such as roads and highways. Apart from the lack of structure
such as clearly defined road edges, the principal difficultly
has to do with large spatial and temporal variations in the
appearance of water in the presence of nearby vegetation and
with reflections from the sky. We propose a self-supervised
method to segment images into “sky”, “river” and “shore” (vegetation + structures). Our approach uses simple assumptions of
geometric structure common to rivers to automatically learn the
relative importance of features such as color, texture and image
location before they are used to segment the image. Our method
performs robustly on three different image sequences from
the Allegheny and Youghiogheny rivers with pixel classification
error rates of less than 5%, outperforming a supervised method,
trained and tested on the same data set. The performance of
the proposed method seems sufficient to allow for guidance of
an autonomous vehicle.
I. I NTRODUCTION
We are interested in a minimal sensor suite based on
passive vision and inertial sensing that could be used to
autonomously explore rivers, mapping their width as it
proceeds. Given sensor pose and camera calibration it is
straightforward to turn segmented images into river maps.
In some cases, the canopy can be so thick and high around
a river so as to block GPS signals and the problem becomes
one of simultaneous localization and mapping. In any case,
a key subproblem to be solved is to automatically segment
visual images such that the extents of the river can be found
accurately. Since rivers can look like roads, one idea is to use
fairly well developed methods that have been used to track
roads and highways with passive vision. Apart from the lack
of structure such as clearly defined road edges, the principal
difficultly of vision-based tracking of rivers has to do with
large spatial and temporal variations in appearance of water
in the presence of nearby vegetation and with reflections
from the sky. See for example, Fig. 1 and Fig. 2.
Such problems are often solved with supervised learning,
an approach where a system is trained based on representative data ahead of time. In our case, supervised learning
would require that we train our river detector with instances
of visual patches that correspond to water in rivers and
Fig. 1. An example image illustrating the variation of river appearance
within a single image.
then use the trained system online. Given the lack of a
representation that would render the appearance of rivers as
a constant even in varying lighting conditions, we find that it
is necessary that any learning be self-supervised. That is, the
system must train itself over time, adapting to local appearance as influenced by terrain, flora and lighting conditions. A
natural question arises in being able to develop such a selfsupervised method– whence the representation? There is a
common intuition that features of color, texture and vertical
location in an image might help. Still, at the outset it is not
clear how to encode/combine different representations.
Here we propose a two step self-supervised method to
perform such segmentation. In the first step we use a simple
heuristic (knowledge of the horizon line) to automatically
determine the relative correlation of between features (a grab
bag of color and texture based features). This allows us to
S. Achar, S. Nuske, S. Scherer and S. Singh are with The Robotics
Institute, Carnegie Mellon University, Pittsburgh, PA 15232, U.S.A.,
{supreeth,nuske,basti,ssingh}@cmu.edu
B. Sankaran is with the GRASP Laboratory, University of Pennsylvania,
Philadelphia, PA, 19104, U.S.A, [email protected]
Fig. 2. This image was taken from almost the same location as Fig. 1 at a
different time of day. The appearance of the river has changed dramatically.
score each image patch on its likelihood of belonging to a
river. Patches with high certainties are used to train a Support
Vector Machine. In the second step, the output of the SVM
segments the image into distinct regions. The advantage of
such an approach is that it assumes no prior to determine
segmentation other than simple geometric assumptions such
as that the river lies below the horizon, and, that river features
are significantly different than other features below the
horizon. All other training of features happens automatically
as often as at every frame.
We have tested our algorithms with image sequences from
two different rivers – the Allegheny and the Youghiogheny
rivers in Pennsylvania, U.S.A.. The image sequences have
distinctly different environmental conditions; sunny and
dusk. We show how supervised learning (training on some
of the sequences and testing on others) by presenting hand
labeled instances of segmentation, provides a reasonable
answer. We show how the results can be improved with our
self-supervised method to the point that in our data sets it
was common to get a misclassification rate around 2% and
never higher than 5%.
In the following sections we review related work (Section
II), show how the features are selected (Section III) and then
present our algorithm (Section IV). The results are presented
in Section V.
II. R ELATED W ORK
The problem of segmenting a river scene from images
can be thought of as a visual classification problem, where
we classify the image into the classes; river and non-river
(shore and sky regions). Classification techniques are categorized into two paradigms: supervised and self-supervised.
Supervised problem setups rely on a labeled training set to
make a global prediction on arbitrary images, while selfsupervised approaches typically operate within one image to
extrapolate information based on a self-generated training set
from another source.
Some examples of supervised appearance classification
and regression to infer scene structure are in Saxena et
al. [2] and Hoiem et al. [3]. These approaches do not
seem applicable to our work, because the appearance (color,
illumination and texture) of a river is highly non-uniform
both spatially across a single image and temporally from day
to day and therefore it is most likely too challenging to learn
a model offline which can encompass the large variety of
appearance that rivers exhibit. Furthermore they are designed
to work in general scenes, whereas we can exploit the general
structure of a river scene to get higher accuracies.
Self-supervised classification, such as the road segmentation work of Dahlkamp et al. [1] collect training examples
online from a portion of the current image (Dahlkamp et al.
use the bottom of the image which is directly in front of the
vehicle) and use to classify the remainder of the image. As is
clear in Fig. 1 we need a different self-supervised approach
for a river, because the areas of the river at the bottom of
the image are vastly different from other areas of the river
in the image.
Some researchers have specifically addressed the problem
of visual water detection using a variety of different passive
vision sensors; non-visible wavelength imagers, polarized
imagery and hyperspectral imaging [4], [5], [6]. For reasons
of cost and weight restrictions we consider a visible light
camera solution.
The work in [7], [8], [9] uses conventional visible passive
imagery to detect bodies of water. In [7] and [8] the authors
detect water by fusing multiple cues like color, brightness
and stereo depth to train a classifier. The recent work in [9]
is for monocular vision and looks to generate a physical
model of bodies of water lying on or nearby roads. They
make the observation that in their application environment
that the color of a water body tends to gradually change
from the leading edge to the trailing edge (a property of
incident angle), when other terrain types do not. They use this
observation to detect water bodies in wide-open areas. Our
application environment is quite different, a river scene that
is not wide-open, variation in color of the river is far more
dependent on the reflections of the surrounding environment
as opposed to incident angle.
Rather than learn or derive a model offline of how the
river will appear, we take an online approach that assumes
a ground plane (river plane) and exploits the knowledge of
the horizon to create self-supervised training sets of features
within the current image.
III. F EATURE S ELECTION
To enable effective river segmentation a discriminative
feature vector (X ∈ Rd ) for describing local appearance is
required. Color is a useful feature for detecting water as
demonstrated in [7] & [8]. Texture is another useful cue,
for example, parts of an image containing the river tend to
be less textured than regions of foliage. Position in an image
provides a strong prior, the higher up in an image a region
is the less likely it is to be part of the river.
We selected features by empirically measuring the performance of different feature vectors describing color and
texture. Each candidate feature vector was used to train a
two class (river and shoreline) linear support vector machine
over a small set of training images and then tested by
using the SVM to segment out the river in images from
a separate evaluation set. Performance of a feature vector
was quantified by the percentage of correctly labeled pixels
averaged over all the images in the evaluation set. All
features were calculated over square 5×5 pixel patches.
For the color descriptor part of the feature vector we used
the RGB, Lab and HSV color spaces individually and in
various combinations to train classifiers. The performance
of these classifiers on labeling river regions in the feature
evaluation test set is show in Table I. A combination of
RGB and Lab colorspaces gave the best performance and
was selected as our color descriptor.
We evaluated the performance of three different texture filters using a similar methodology. The first was the response
to a set of Laws’ masks [10] over 4 different scales of the 3
channels of the Lab image. We used 8 of the 9 3×3 Laws’
Masks leaving out the mask that performs low pass filtering,
so the resulting texture descriptor had length 4 × 3 × 8 = 96.
The second texture descriptor considered was the DAISY
descriptor [11] without normalization. The third alternative
was a bank of Gabor filters with 4 scales and 6 orientations.
There was no significant difference in terms of performance
between the three filter sets (Table I) so we chose the Laws’
masks because it was the fastest to compute.
TABLE I
P ERFORMANCE OF D IFFERENT F EATURE V ECTORS FOR S EGMENTATION
INTO T WO C LASSES
Feature
RGB
Lab
HSV
RGB+HSV
RGB+Lab
RGB+Lab+HSV
Color
Dimensionality
3
3
3
6
6
9
Error Rate
47.05%
35.61%
46.62%
44.43%
32.90%
37.60%
Texture
Laws’ Masks
DAISY
Gabor
96
200
24
26.81%
26.76%
27.19%
Combined
Color+Texture
Color+Texture+Height
102
103
19.73%
4.61%
When combined, color and texture features perform better
than either cue alone. Image position, specifically the height
of a region in image coordinates, provides a strong contextual
cue for segmentation and hence is included in the feature
vector. The bottom of Table I shows the labeling error rate
for the combined color and texture descriptor and for our
final choice of feature descriptor on the feature evaluation
set.
IV. S ELF - SUPERVISED R IVER S EGMENTATION
Riverine environments vary widely in appearance which
makes building a single classifier that works reliably in
different conditions difficult. Such a classifier would require
labeled training images captured in a variety of environments
and under different conditions. Its performance in an environment dissimilar to those seen earlier would not be assured.
Instead of attempting to learn offline a universal model of
river appearance, we automatically learn a new appearance
model online for each image presented to the algorithm.
To make this possible we utilize knowledge about the
position of the horizon in each image and assume that
anything appearing above the horizon line is not part of the
river. In our application domain, the inertial measurement
unit and altimeter on the air vehicle would be used to
calculate the horizon line. In the absence of sensor data, a
vision based estimate of the horizon can be used instead [12].
The assumption that the river can not be above the horizon is
valid except for situations in which the river has a significant
upward slope (these are unusual scenarios in which we do
not expect to operate like upstream on a rapid or at the
base of a waterfall). An additional assumption we make is
that the appearance of the above horizon regions is fairly
representative of the appearance of ther entire shore area in
that image.
As a preprocessing step, we first segment out the sky and
ignore it during all further computations. To build an online
appearance model for discriminating between river and shore
regions, we automatically extract two types of patches from
the image, those that are part of the shore region and those
that are likely to be part of the river. Generating shore patches
is trivial, we just sample patches lying above the horizon line.
We find the patches below the horizon line that are most
dissimilar in appearance to those above the horizon and use
them as candidate river patches. We then use these patches
as training examples for a linear SVM classifier which is
then used to classify all parts of the image. Details of each
step in the algorithm are given below.
A. Feature Extraction
As described in the Section III, the image is divided into
square patches and a feature vector (X ∈ Rn ) is computed for
each patch which describes its color, texture and position.
B. Sky Detection
A linear support vector machine trained over a set of
labeled images is used to detect the sky region in each image.
The sky is relatively easy to segment out with a globally
trained classifier as its appearance is not as variable as that
of the river. Discarding the sky before further processing
reduces computational effort and prevents confusion with
shiny, mirror like parts of the water’s surface which often
have an appearance similar to that of the sky. Fig. 3(a) shows
an example of sky detection. The remainder of the image
needs to be classified as either being part of the river or part
of the shore.
C. Appearance Modeling
The goal of the appearance modeler is to assign to each
patch a probability of being part of the river (R) or being
part of the shore (¬R) on the basis of the feature vector (X)
that describes it. By Bayes’ Rule we have
P(¬R | X) =
P(X | ¬R)P(¬R)
P(X)
(1)
Since the labels (R and ¬R) are what we are trying to
determine, P(X | ¬R) can not be calculated directly. We
define a new boolean variable H which is true for regions
below the horizon line and false for regions above the
horizon. The value of H for each position in the image
is known because the horizon is known. We assume that
the appearance of the shore regions above the horizon is
representative of the entire shore region in the image, this
implies that P(X | ¬R) ≈ P(X | ¬H) which gives
P(¬R | X) ≈
=
P(X | ¬H)P(¬R)
(2)
P(X)
P(X | ¬H)P(¬R)
(3)
P(X | ¬H)P(¬H) + P(X | H)P(H)
The appearance distributions P(X | ¬H) and P(X | H) are
modeled using Chow Liu trees [13]. A Chow Liu tree is a
method for approximating the joint probability distribution
of a set of discrete random variables by factorizing it into
a product of second order distributions. A Chow Liu tree is
optimal in the sense that it is the distribution of its class that
minimizes the KL divergence to the real distribution.
The feature vector X contains color, texture and image position information and has high dimensionality as it contains
many texture filter responses. For the Chow Liu appearance
modeling an abridged feature vector X̂ ∈ Rd is created in
which the texture descriptor subvector is replaced by its L2
norm because using the full length feature vector would slow
down computation. Additionally, X̂ does not include image
position. Since we are modeling only appearance including
position information in X̂ would introduce an undesirable
bias.
The features in feature vector X̂ are continuous valued
while Chow Liu trees work over discrete valued random
variables. To solve this problem, X̂ ∈ Rd for each image
patch is converted into a discretized feature vector X̃ ∈ Sd
where each feature in X̃ is assigned to 1 of 16 levels
(S = {0, 1, 2, . . . , 15}) using equally sized bins that span the
range of values taken by that feature.
Two Chow Liu trees are built, one using the the feature
vectors of patches above the horizon (X̃ ∈
/ H) to model
P(X | ¬ H) and another using the feature vectors of
below horizon patches (X̃ ∈ H) to model P(X | H). We use
subscripts to denote individual features in a feature vector
so X̃i is the ith feature in feature vector X̃. A Chow Liu tree
is a maximal spanning tree of the mutual information graph.
The mutual information graph has one node for each feature
(X̃i ) and an edge between every pair of nodes (X̃i and X̃ j )
which is weighted by their mutual information I(X̃i ; X̃ j ).
|S|−1 |S|−1
I(X̃i ; X̃ j ) =
∑ ∑
xi =0 x j =0
p(xi , x j ) log
p(xi , x j )
p(xi )p(x j )
(4)
Where p(xi ) is the probability that the ith feature in
X̃ takes on the value xi , P(X̃i = xi ). Similarly, p(x j ) is
P(X̃ j = x j ) and p(xi , x j ) is the joint probability distribution
P(X̃i = xi , X̃ j = x j ). All these distributions are estimated
(a) P(¬R | X)
directly from X̃ ∈
/ H and X̃ ∈ H by counting feature value
occurrences and co-occurrences. Edges are deleted from the
mutual information graph to form a maximal spanning tree.
The resulting tree encodes the approximate distribution as a
product of pairwise conditionals. We can calculate P(X̃ = x̃),
the probability of occurrence of a particular feature vector
x̃ ∈ Sd = {x̃1 , x̃2 , x̃3 , . . . , x̃d } as follows
P(X̃ = x̃) ≈ ∏ p(X̃i = x̃i | X̃π(i) = x̃π(i) )
(5)
where π(i) is the parent of node X̃i in the tree. The two
Chow Liu trees for P(X | ¬H) and P(X | H) are used in (2) to
calculate P(¬R | X) for each patch in an image. An example
is shown in Fig. 3(a).
D. Classifier Training
For each patch, the probability of being part of the shore
region P(¬R | X) is calculated using (2). The patches that are
least likely to be on the shore (P(¬R | X) < T ) are used as
candidate river patches. Using a low value for T reduces the
chance that shore patches are accidentally used as candidate
river patches, but if T is set too low then the selected region
will be too small to provide a good representation of the river
region. In our experiments, T was set to 0.01. The shore
patches above the horizon and the candidate river patches
with P(¬R | X) < T are used to train a two class (river/shore)
linear support vector machine. Fig. 3(b) shows the selected
river and shore training examples for an input image.
The SVM uses the unabridged, continuous valued feature
vectors X, including image position and the complete texture
descriptor. The support vector machine is trained using an
online stochastic gradient descent method. Since we are
learning a new classifier for each new frame, while the
appearance of the river is likely to remain fairly consistent
over short periods of time, we initialize the SVM training
using the SVM model learnt on the previous frame. This
reduces the number of gradient descent iterations needed to
train the classifier for each new frame. Fig. 3(b) shows the
river and shore training examples picked by the algorithm in
an image.
(b) Training Examples
(c) Final Result
Fig. 3. Steps in the Self-Supervised River Segmentation Algorithm: (a) P(¬R | X): warm colors are used for values closer to 1 (probably shore region).
The detected sky region is marked in black (b) Training examples: The algorithm finds shore regions (marked in green) and candidate river regions (marked
in red) for training a classifier. (c) Final result: The classifier is run on the entire image to classify each patch. The detected extent of the river is marked
in red.
E. Detecting Water
The learnt SVM model is used to classify all the patches
in the image. As a post processing step, small holes in
the labeling are filled using morphological operators. An
example final segmentation result is shown in Fig 3(c)
approach works well when used on datasets it had been
trained on (the self column) but performance degrades when
run on the other two datasets that were not seen during
training (Leave one out). This suggests that a supervised
approach does not generalize well to new environments.
TABLE II
S UPERVISED S EGMENTATION P ERFORMANCE : M EAN E RROR (S TD DEV )
V. R ESULTS
A. Datasets and Groundtruth
Three datasets were used for evaluating our algorithm.
The data collection setup used was a tripod mounted camera
(a SANYO VPC-CG102BK for the Allegheny Day dataset
discussed following and a Canon SX200IS for the other two
datasets) placed onboard a small motorboat. The cameras
captured 1280×720 images which were downsampled to
640×360 for processing. Two of the datasets were collected on the Allegheny river at Washington’s Landing
in Pittsburgh, PA, U.S.A, the third was collected on the
Youghiogheny at Perryopolis, PA, U.S.A.. One of the Allegheny datasets (referred to as Allegheny Day) was collected
during the afternoon and the other (Allegheny Dusk) was
collected in the evening close to dusk. The dataset on
the Youghiogheny was collected around noon. Each dataset
contained between 120 and 150 images that were manually
labeled to provide groundtruth for performance evaluation.
The labels used were river, shore (vegetation, structures etc.)
and sky. An example ground truth labeling is shown in Fig. 4.
The performance metric used in all tests and while reporting results is the percentage of pixels misclassified by
the algorithm when compared against ground truth. Since
the classifiers used generate only two output labels and we
are not interested in differentiating between the shore region
and sky we treat them as a both as a single class (non-river)
during performance evaluation.
Allegheny Day
Allegheny Dusk
Youghiogheny
Self
LOO
3.77% (0.64%)
3.55% (0.46%)
3.50% (0.34%)
4.01% (0.63%)
8.13% (1.53%)
4.47% (0.41%)
C. Self-supervised Segmentation Results
We evaluated the self-supervised segmentation algorithm
described in Section IV on the three datasets. Error rates
are shown in Table III. On all three datasets, the self
supervised algorithm outperformed a supervised classifier
that was tested on previously unseen datasets. Even when
the supervised algorithm was trained on a subset of images
from the same dataset it was tested on, it only outperformed
the self supervised method on Allegheny Dusk. This is
significant considering how the self-supervised method uses
no manually labeled input.
Fig. 5 shows some example images from the three datasets
and the detected extent of the river along with the major
intermediate outputs of the algorithm. Some images on which
our algorithm failed to detect the river correctly are shown in
Fig. 6. In Fig. 6(a) part of the tree was misclassified because
it was in the shadows and also parts of the background just
above the horizon were incorrectly labeled as parts of the
river. In Fig. 6(b) a large swath of the grass on the bank was
marked as part of the river because a few grass patches were
selected as positive river examples for the classifier.
TABLE III
S ELF S UPERVISED S EGMENTATION P ERFORMANCE
Location
(a)
(b)
Fig. 4. (a) an image from the Allegheny Day dataset and (b) its groundtruth
labeling with the river marked in red, shore (foliage/structures etc.) colored
green and the sky in blue.
B. Supervised Segmentation Results
We investigated how well a supervised classifier would
be able to generalize to images from previously unseen
environments. For each dataset two classifiers were built,
the first classifier (Self in Table II) was trained using two
randomly selected images from the same dataset and the
second classifier (LOO: Leave one out) was trained on two
images selected from the other two datasets. The error rates
for these classifiers were averaged over 16 trials and are
tabulated in Table II. It can be seen that the supervised
Allegheny Day
Allegheny Dusk
Youghiogheny
Mean Error Rate
2.71%
4.15%
2.67%
VI. C ONCLUSIONS AND F UTURE W ORK
We presented a method for using monocular vision to
estimate the extent of a river in an image. We demonstrated
that a naive supervised technique is unlikely to generalize
well to different environments. We formulated and evaluated
a self-supervised alternative that automatically generates a
model of river appearance when presented with an image by
leveraging assumptions that can be made about the structure
of the environment and the horizon.
Our current system is completely memoryless as it builds
a new river appearance model for every input image. Using some history of previous river appearance to build an
(a) Allegheny Day
(b) Allegheny Day
(c) Allegheny Day
(d) Allegheny Dusk
(e) Allegheny Dusk
(f) Youghiogheny
(g) Youghiogheny
(h) Youghiogheny
Fig. 5. Example frames from different datasets on which the self supervised algorithm performed well. For each example the top left quadrant is the
input, the top right quadrant is a heatmap showing the probability of a patch being part of the shore based on its appearance (the regions detected as part
of the sky are marked black), the bottom left quadrant shows the automatically detected river and shore regions for training a classifier and the bottom
right quadrant shows the extent of the river as determined by the algorithm.
adaptive model could increase performance and provide
robustness against failure. The most serious failure mode in
our algorithm is when novel objects (boats, piers etc) appear
below the horizon but do not get marked as non-river. Our
algorithm is unable to handle this adequately for two reasons.
Firstly it does not maintain a model of river appearance and
secondly it assumes that the appearance of the above horizon
region adequately models all non-river objects. Keeping a
model of river appearance learnt from previously seen images
would allow the algorithm to detect novel non-river objects
below the horizon.
We would also like to investigate the use of priors on likely
river shapes and performing inference over multiple frames
in a video sequence to increase accuracy.
R EFERENCES
[1] H. Dahlkamp, A. Kaehler, D. Stavens, S. . Thrun, and G. . Bradski,
“Self-supervised monocular road detection in desert terrain,” in Proc.
Robotics Science and Systems Conference (RSS’06), Philadelphia, PA,
Aug. 2006.
[2] A. Saxena, M. Sun, and A. Ng, “Learning 3-d scene structure from
a single still image,” Computer Vision, 2007. ICCV 2007. IEEE 11th
International Conference on, pp. 1–8, 2007.
[3] D. Hoiem, A. A. Efros, and M. Hebert, “Automatic photo pop-up,”
Proceedings of ACM SIGGRAPH 2005, Jan 2005.
(a) Allegheny Dusk
[4] A. Sarwal, J. Nett, and D. Simon, “Detection of small water bodies,”
in PrecepTek Robotics Technical Report (ADA433004’04), Littleton,
CO, 2004.
[5] B. Xie, Z. Xiang, H.Pan, and J.Liu, “Polarization-based water hazard
detection for autonomous off-road navigation,” in Proc. IEEE International Conference on Intelligent Robots and Systems (IROS’07), San
Diego, CA, Sept. 2007.
[6] H. Kwon, D. Rosario, N. Gupta, M. Thielke, D. Smith, P. Rauss,
P. Gillespie, and N. Nasrabadi, “Hyperspectral imaging and obstacle
detection for robotics navigation,” U.S Army Research Laboratory
Technical Report, Tech. Rep. (ARL-TR-3639’05), Sept. 2005.
[7] A. Rankin, L. Matthies, and A. Huertas, “Daytime water detection
by fusing multiple cues for autonomous off-road navigation,” in Proc.
24th Army Science Conference (ASC’04), Orlando, FL, Nov. 2004.
[8] A. Rankin and L. Matthies, “Daytime water detection and localization
for unmanned ground vehicle autonomous navigation,” in Proc. 25th
Army Science Conference (ASC’06), Orlando, FL, Nov. 2006.
[9] ——, “Daytime water detection based on color variation,” in
Proc.IEEE International Conference on Intelligent Robots and Systems
(IROS’10), Taipei, Taiwan, Oct. 2010.
[10] K. Laws, “Rapid texture identification,” SPIE, pp. 376–380, 1980.
[11] E. Tola, V. Lepetit, and P. Fua, “A fast local descriptor for dense
matching,” in Conference on Computer Vision and Pattern Recognition, Alaska, USA, 2008.
[12] S. Ettinger, M. Nechyba, P. Ifju, and M. Waszak, “Towards flight
autonomy: Vision-based horizon detection for micro air vehicles,”
Proceedings of the IEEE/RSJ International Conference on Intelligent
Robots and Systems, Jan 2002.
[13] C. Chow and C. Liu, “Approximating discrete probability distributions
with dependence trees,” Information Theory, IEEE Transactions on,
vol. 14, no. 3, pp. 462 – 467, may. 1968.
(b) Youghiogheny
Fig. 6. Some images on which the self supervised algorithm failed. In (a) the base of the tree and some regions just above the horizon line are marked
as part of the river. In (b) the grass on the bank is marked as part of the river because a few patches on the bank were picked as river training examples.